Louis Vervoort's Problem Solving in Philosophy: How to Do Philosophy in the Age of Ultra-Intelligent AI (springer, 2026), raises a question that professional philosophers might prefer not to contemplate. What happens when artificial intelligence becomes better at philosophy than philosophers?
Vervoort's answer is that philosophy itself needs to change. Instead of philosophers spending centuries arguing over questions that never seem to get settled, philosophical problems should increasingly be approached as scientists approach competing theories. Construct possible solutions, compare how much they explain, look for contradictions, assess their coherence and choose the theory that performs best. Vervoort calls his preferred goal a "maximally coherent theory." Once sufficiently powerful AI enters the picture, machines could search vastly more arguments, objections and possible theories than any human philosopher could hope to master.
There is something attractive about this. Philosophy certainly has a problem with problems that never die. Scientists disagree, but experiments can eventually eliminate possibilities. Mathematicians may spend centuries on a problem, but once a valid proof appears there is a recognisable sense in which the problem has been solved. Philosophers can still be arguing about essentially the same questions after 2,000 years.
AI also presents an obvious challenge to academic philosophy. Imagine an artificial intelligence capable of reading virtually the entire philosophical literature, remembering every objection ever published, finding contradictions that generations of scholars missed and generating millions of possible combinations of existing positions. Such a machine would possess enormous advantages over the professor sitting in an office with a few shelves of books and a human lifetime in which to read them.
So let us give Vervoort the strongest possible case. Forget today's hallucinating error-prone chatbots. Imagine something vastly more powerful, backed by the resources of one of the great AI companies. Thousands or tens of thousands of specialised AI agents work simultaneously on one philosophical problem. Some reconstruct every argument ever offered. Others search for objections. Others defend those arguments against the objections. Still others act as hostile referees trying to destroy every proposed solution. Higher-level systems integrate the surviving arguments and begin the process again.
What problem should we give them?
Free will would be tempting, but there is an obvious difficulty. Much of the argument concerns what we mean by "free will." God would present similar problems concerning definitions, evidence and metaphysical assumptions.
There is a better test: the problem of the criterion.
It is one of the deepest problems in epistemology, the branch of philosophy concerned with knowledge. Put simply, it asks how we distinguish truth from error, and how we justify that.
Suppose I claim to possess a criterion that tells me which beliefs are true. You reasonably ask how I know that my criterion is reliable. I might answer that I tested it against beliefs that I already know to be true.
But now the trap closes. How did I know those beliefs were true before I possessed a reliable criterion for identifying truth?
We seem to need known truths in order to establish the correct criterion, while at the same time needing the correct criterion in order to establish which beliefs are known truths. The problem is ancient in spirit and appears throughout the sceptical tradition, as a key argument to show that knowledge does not exist. In modern philosophy it became especially associated with what Roderick Chisholm called the "problem of the criterion."
This seems an ideal challenge for Vervoort's vision.
Imagine an AI research programme called Project Criterion. Give it everything ever written about scepticism, truth, knowledge, evidence and epistemic justification. Let specialised agents work through Sextus Empiricus, Descartes, Hume, Kant, Peirce and the modern literature. Let other teams investigate foundationalism, coherentism, reliabilism, Bayesian epistemology and every serious alternative; it does not matter here what these terms mean, they are just theories of knowledge. Tell the system not merely to summarise this literature but to start again from square one.
Its instructions would be simple in principle: reconstruct every serious solution in its strongest possible form. Find every hidden assumption. Generate new solutions that no human philosopher has previously considered. Subject every proposed answer to relentless adversarial attack. Determine whether there is a non-question-begging criterion by which true beliefs can be distinguished from false ones. If there is, demonstrate it. If there is not, explain why not.
Now give Project Criterion six months, enormous computing resources and billions or trillions of tokens.
This would be a genuine test of the claim that philosophical problems persist largely because human beings have lacked sufficient intellectual power to solve them.
Perhaps Vervoort would be vindicated spectacularly. The AI might discover that philosophers have spent 2,000 years trapped inside a conceptual framework containing an unnoticed mistake. It might produce a new solution that is neither foundationalist nor coherentist nor sceptical, because those familiar alternatives turn out to divide the conceptual territory incorrectly. Human philosophers might read the resulting argument and realise, with some embarrassment, that the problem had finally been cracked.
If that happened, we should take the prospect of machine philosophy extremely seriously.
But there are other possible outcomes. Project Criterion might conclude that no proposed criterion can be justified without presupposing at least some epistemic principles whose reliability cannot themselves be established without circularity. That would also be an important result, although it would hardly amount to philosophy becoming theoretical physics.
There is an even more interesting possibility. The AI might announce that Theory A provides the best solution because it achieves the greatest coherence, explanatory power and coverage with the fewest contradictions.
At that point the old sceptic raises his hand.
"How do you know those are the correct criteria?"
That question goes directly to the difficulty in Vervoort's project.
Suppose our super-AI ranks philosophical theories according to coherence, explanatory reach, simplicity, consistency and ability to answer well-posed questions. Those sound like perfectly reasonable intellectual virtues. But why those virtues? How should they be weighted? Is consistency always more important than explanatory reach? Is one contradiction fatal? Should simplicity beat comprehensiveness? How much evidence should be required before explanatory elegance counts for anything?
Who designed the scoreboard?
If the AI replies that coherence is important because coherent theories are more likely to be true, the sceptic can ask how we established that proposition. If it appeals to the historical success of coherent theories, questions about induction immediately return. If it appeals to explanatory power, we can ask why explanatory power reliably tracks truth. If it appeals to empirical success, we can ask what criterion establishes that successful prediction warrants belief in the truth of the underlying theory. The problem of the criterion has returned inside the machinery supposedly designed to solve it.
This is why making philosophy quantitative does not automatically make it scientific. Giving philosophical virtues numerical values may clarify our reasoning enormously. It may expose inconsistencies and force philosophers to reveal assumptions they previously left hidden. That would be valuable work. But assigning numbers to philosophical criteria does not establish that we selected the correct criteria in the first place. The method risks becoming philosophy pretending that it has escaped philosophy.
This is also where comparison with mathematics and theoretical physics becomes revealing. In a tightly specified mathematical problem, the rules have largely been settled before the machine begins. Definitions have been supplied, the formal framework established and standards of proof agreed upon. An AI that produces a valid proof has accomplished something extraordinary, but it was given a relatively clear target.
Project Criterion receives no such luxury because determining what counts as a satisfactory solution is partly what the problem is about. That is why the problem of the criterion provides a harder test of Vervoort's programme than simply asking AI whether human beings possess free will. We would be asking AI to justify the intellectual machinery by which it proposes to tell us which philosophical answers are correct.
There is an analogy here with Hume's famous problem of induction. Science relies upon reasoning from observed cases to unobserved cases. We expect tomorrow to resemble yesterday sufficiently for prediction to be possible. Vervoort's response is essentially to allow an inductive principle to function as a basic axiom. Science needs it, so we stop demanding an impossible further proof.
That certainly allows science to proceed. But science needing induction is not the same thing as philosophy justifying induction. Hume can still ask why we are epistemically entitled to accept the axiom. Declaring the question a pseudo-problem does not necessarily answer it. It may simply announce that our system has decided to stop asking.
The problem of the criterion generalises this difficulty. Every philosophical system eventually needs standards by which arguments, evidence and conclusions are assessed. If those standards themselves become objects of philosophical dispute, moving them into the machinery of an AI does not make the dispute disappear.
Yet this does not mean the proposed AI experiment would be pointless. Quite the opposite. I would love to see it done.
Take the sort of enormous resources now being devoted to making AI systems better programmers, mathematicians and scientific researchers and divert a serious portion of them to one fundamental philosophical problem. Do not ask the machine for another literature review. Do not ask it to imitate Kant or Hume. Do not tell it which philosophical school is correct. Do not even assume that the traditional formulation of the problem is right.
Tell thousands of agents to start from the ground up. Read everything. Attack everything. Assume as little as possible. Generate conceptual possibilities human philosophers have never considered. And do not come back until you have either solved the problem of the criterion or identified precisely why it cannot be solved.
There would be several fascinating possible results. The AI might produce a genuinely new solution that human specialists cannot defeat. It might demonstrate that the problem is insoluble under assumptions shared by all existing approaches. It might discover that the traditional problem itself contains a hidden conceptual error. Or it might return with an immense map showing that every route eventually reaches some proposition that must be accepted without the kind of independent justification originally demanded. Any of those outcomes would teach us something.
But suppose after six months Project Criterion announces: "We have evaluated every serious theory. Theory 8,472,119 is the maximally coherent solution."
The philosopher asks, "By what criterion?"
The machine supplies its criteria.
The philosopher asks, "How did you establish that those are the correct criteria?"
If the machine answers by appealing to another set of criteria, we have simply moved upstairs. If it ultimately reaches an axiom that cannot itself be independently justified, the original difficulty has not vanished. We have located it with unprecedented precision.
This also reveals something important about claims that AI will make philosophy obsolete. Perhaps AI really will make a considerable amount of professional philosophy obsolete. A large amount of academic work consists of searching existing literature, combining positions, finding objections, inventing distinctions and producing another variation upon familiar theories. Machines may eventually perform such work much better than human beings. That would change the sociology of philosophy. It would not necessarily abolish philosophy.
Indeed, if Project Criterion discovered an extraordinary solution, we might conclude not that AI had destroyed philosophy but that AI had become an extraordinarily powerful philosopher. It would have done what philosophers have always attempted to do: question assumptions, construct arguments, find contradictions, invent concepts and push inquiry beyond the limits reached by its predecessors.
The real test would come if its reasoning surpassed human comprehension. Suppose Project Criterion announces that it has solved the problem but its proof requires conceptual structures no human being can understand. Should we believe it?
Immediately another philosophical problem appears. What justifies trusting the conclusion of an intelligence whose reasoning we cannot evaluate? We are back to criteria again.
This is why Vervoort's vision deserves to be taken seriously without accepting its strongest conclusion. AI could transform philosophy. It could expose bad arguments, discover connections no human scholar could see, explore millions of possible theories and perhaps solve problems we presently regard as permanent. It may become extraordinarily good at playing the philosophical game. But before allowing the machine to declare the winner, we still need to know why these are the right rules for the game.
So here is a challenge for the coming age of ultra-intelligent AI. Before asking 30,000 AI agents which philosophical theory is true, give them the more fundamental problem philosophers have never satisfactorily settled:
What criterion entitles us to call anything true in the first place?
Give them six months and all the computing power they want.
I genuinely want to hear the answer. My challenge: let's see them test this one!
https://www.youtube.com/watch?v=RmvQtxc9R3Q
https://link.springer.com/content/pdf/10.1007/978-3-032-17756-8.pdf